A new research report released by Morgan Stanley analyst Howard Kao on Friday (22nd) points out that the cost of AI infrastructure is experiencing a structural surge. Nvidia's upcoming next-generation Vera Rubin AI server rack, sourced from original design manufacturers (ODMs), costs approximately $7.8 million, almost double the current GB300 Blackwell (less than $4 million). This near-doubling increase per generation reflects a dramatic shift in the value distribution of the AI ??hardware supply chain, not just due to rising core GPU prices.
It's worth noting that although Nvidia's stock fell nearly 2% the day after its earnings report on Wednesday, memory-related stocks surged 6%-10%, confirming the core logic of the Morgan Stanley report: the biggest beneficiaries of this round of price increases are spreading from core computing units to peripheral key components.
Morgan Stanley's research on the bill of materials (BOM) indicates that memory has risen from a secondary role to a major cost driver. In the older GB200 system, memory accounted for only about 5% to 10% of rack material costs. However, in the VR200 system, benefiting from increased capacity and rising prices, this percentage has skyrocketed to 25% to 30%, representing an absolute cost increase of 435%, directly compressing the GPU's share (from approximately 65% ??to 51%).
Morgan Stanley points out that this wave of price increases is comprehensive. Printed circuit boards (PCBs) have seen a 233% cost surge due to the introduction of new modules and layer upgrades; multilayer ceramic capacitors (MLCCs) have seen a 182% increase in cost due to increased demand for new components; ABF substrates have increased by 82% due to the doubling of chip count; and power supplies and liquid cooling components have also increased by 32% and 12%, respectively.
However, analysts also point out a key variable: if hyperscale cloud providers such as Microsoft and Google choose to bypass Nvidia and directly purchase SOCAMM memory modules, rack prices could drop to approximately $6.7 million. This would directly impact Nvidia's memory resale revenue, a key focus for subsequent market monitoring.
For ODM (Original Design Manufacturer) companies responsible for assembly, the Morgan Stanley report breaks the market consensus that standardization will compress added value. According to Morgan Stanley's estimates, the ODM value-added portion of Rubin racks will actually increase by 35% to 40%, rising from approximately US$108,000 for GB300 to approximately US$149,600. Although the gross margin may slightly decrease from 2.7% to 1.9%, the increase in absolute profit is more crucial.
However, the supply chain model is quietly shifting. Foxconn, Quanta, and others have successively mentioned the "consignment model," where cloud vendors purchase core components themselves, and ODMs are only responsible for assembly. While this can alleviate the working capital pressure on ODMs, it may compress revenue scale in the long run.
Furthermore, power supply specifications are evolving towards high-voltage direct current (HVDC). Rubin Ultra is expected to adopt 800V DC on a large scale by 2027, and major power supply manufacturers such as Delta Electronics have already actively cooperated in this area.
Overall, AI server racks are entering a new phase of high value and high complexity, and the redistribution of the supply chain profit pool has only just begun.
Tags：